XTalker (Xeon SadTalker) is a faster and optimized implementation of SadTalker. It utilizes low precision and parallelism to boost the inference speed by up to 10x compared to the original implementation on one Sapphire Rapids (SPR) Xeon CPU (without any GPU used). We now have optimized both the rendering and the enhancing stages, which are the main time-consuming stages for the original SadTalker. This project is experimental and actively under development. Welcome to any advices and ideas.

• Install PyTorch

pip install torch==2.0.0+cpu torchvision==0.15.1+cpu torchaudio==2.0.1 --index-url --index-url https://download.pytorch.org/whl/cpu

• Install related dependencies of SadTalker following README

• Install Intel Extension For PyTorch

python3.8 inference.py --driven_audio xxx.wav --source_image xxx.jpg --result_dir ./results --cpu --bf16

Normally IOMP is referred to the Intel openmp library. Here I just use this word to refer to my parallelled implementation of SadTalker on Xeon CPU and together with the libiomp.so preloaded.

• (Optional) Preload the optional library to get an optimal speedup

export LD_PRELOAD=<PATH TO tcmalloc.so>:<PATH TO libiomp5.so>

• Generate the parallelized execution script based on your hardware

Without enhancer

python generate_distributed_infer.py --slot=<parallelism number> --driven_audio xxx.wav --source_image xxx.jpg

With enhancer

python generate_distributed_infer.py --slot=<parallelism number> --driven_audio xxx.wav --source_image xxx.jpg --enhancer gfpgan

Please change the parallelism number with your expected parallelism on your CPU.

IOMP can be used together with IPEX bf16 to get a further acceleration. You just need to add --bf16 when generating the run_distributed_infer_xx.sh, like following

Without enhancer

python generate_distributed_infer.py --slot=<parallelism number> --driven_audio xxx.wav --source_image xxx.jpg --bf16

With enhancer

python generate_distributed_infer.py --slot=<parallelism number> --driven_audio xxx.wav --source_image xxx.jpg --bf16 --enhancer gfpgan

• Run the script

bash run_distributed_infer_<parallelism number>.sh

Go to branch int8

pip install neural-compressor

python3.8 inference.py --driven_audio xxx.wav --source_image xxx.jpg --result_dir ./results  --cpu

For the first time the model will be quantized and saved to generator_int8, and the following runs will load the quantized model from generator_int8 to do inference.

IOMP can be used together with int8 quantization (powered by INC) to get a further acceleration.

Go to branch int8

pip install neural-compressor

python generate_distributed_infer.py --slot=<parallelism number>

The recommended parallelism number to take is to increase from 4 to 8, 16...

Again, remember to run at least twice the following command, when the first run (take a bit longer) is to dump the int8 model, and then in following runs (faster) we can just do inference.

bash run_distributed_infer_<parallelism number>.sh

int8 + IOMP still have some accuracy loss (maybe because of insufficient calibration or int8 precision itself), which is mainly manifested by noises in the generated frames. This is still under developing.

Based on Discussion #457, PIRender is proved to be faster than facevid2vid for the face rendering stage, and it is orthogonal to our optimization. We integrate this feature into xtalker and apply our optimization to it to get further speedup. To use it, you can simply add a parameter --facerender pirender to the above commands.

python generate_distributed_infer.py --slot=<parallelism number> --driven_audio=xxx.wav --source_image=xxx.jpg --facerender=pirender
bash run_distributed_infer_<parallelism number>.sh

Notice: PIRender model now has accuracy issue in bf16 optimization so we disable bf16 here. If you want to involve this model, please also follow PIRender's license for appropriate usage.

You can at any time do "garbage collection" when there is an abnormal exit by the following command:

bash gc.sh

We compare the end-to-end inference time (in seconds) between XTalker on SPR and SadTalker with one A100 40GB. The input audio is a 20-second audio file and an image and the output is a 20-second video file. The figures show that our best implementation (IOMP + bf16) is 1.33x faster than A100 with almost no accuracy loss.

Notice: All SPR experiments are ran with the libiomp.so preloaded.

XTalker is adapted from SadTalker. We thank the related authors for their great work!

This is not to be used for any commercial purposes and is not an official project from Intel. If you have any problems, please contact sihan.chen@intel.com.